Date:	Mon, 9 Jun 2008 15:41:41 +0200
From:	Laurent Pinchart <laurentp@...-semaphore.com>
To:	Ben Dooks <ben-linux@...ff.org>, g@...nity.fluff.org
Cc:	Ben Dooks <ben-linux@...ff.org>, netdev@...r.kernel.org
Subject: Re: DM9000 issue with mem resource management

Hi Ben,

On Saturday 07 June 2008 23:35, Ben Dooks, g@...nity.fluff.org wrote:
> On Fri, Jun 06, 2008 at 11:01:42AM +0200, Laurent Pinchart wrote:
> > Hi Ben,
> > 
> > On Thursday 05 June 2008 19:40, Ben Dooks wrote:
> > > On Thu, Jun 05, 2008 at 05:42:58PM +0200, Laurent Pinchart wrote:
> > > > Hi everybody,
> > > > 
> > > > I ran into a resource-related bug in the DM9000 driver.
> > > > 
> > > > When the platform device has only 2 resources, dm9000_probe() doesn't
> > > > set db->irq_res, which results in a segfault when the pointer gets
> > > > dereferenced in dm9000_open().
> > > > 
> > > > I tried to fix the issue, and found out that the resource management
> > > > code is quite broken.
> > > 
> > > Personally, I'm thinking about just removing the case for 2, and making
> > > it three resources only. The following in-kernel machines use the
> > > following resources:
> > > 
> > > arch/arm/mach-at91/board-sam9261ek.c			3
> > > arch/arm/mach-pxa/cm-x270.c				3
> > > arch/arm/mach-pxa/em-x270.c				3
> > > arch/arm/mach-pxa/trizeps4.c				3
> > > arch/arm/mach-pxa/colibri.c				3
> > > arch/arm/mach-s3c2410/mach-bast.c			3
> > > arch/arm/mach-s3c2410/mach-vr1000.c			3
> > > arch/blackfin/mach-bf527/boards/ezkit.c			2
> > > arch/blackfin/mach-bf533/boards/H8606.c			2
> > > arch/blackfin/mach-bf533/boards/ip0x.c			3
> > > arch/blackfin/mach-bf537/boards/generic_board.c		2
> > > arch/blackfin/mach-bf537/boards/stamp.c			2
> > > 
> > > As you can see, the #3 outweigh the #2. Unless anyone else objects, I
> > > will add a patch to reduce this to the case where the driver expects
> > > 3 resources, and ask the users of #2 to submit changes for their
> > > bots.
> > >  
> > > > If I understand things correctly, specifying 3 resources makes the
> > > > DM9000 driver ioremap() the memory, while specifying 2 resources
> > > > implies that the platform code already ioremap()ed the memory. Is that
> > > > right ? 
> > > > 
> > > > If so, why does dm9000_probe() call request_mem_region() on
> > > > ioremap()ed memory ?
> > > 
> > > Hmm, dunno. I really have no idea why this happened. I don't think this
> > > was my fault!
> > >  
> > > > Wouldn't it also be simpler to use release_mem_region() in 
> > > > dm9000_release_board() instead of release_resource() + kfree() ?
> > > 
> > > If you already have the resource, this is a reasonably simple way of
> > > ensuring you're disposing of the right resource.
> > >  
> > > > I'd be grateful if someone could confirm my assumptions. I'll then
> > > > submit a patch to fix those issues.
> > > 
> > > My recommendation is that we remove the 2 case completely, so I'd not
> > > bother trying to fix this.
> > 
> > I'm not too sure on that one.
> 
> Well, having thought about it some more, the 2 address case is also an
> abuse of the IORESOURCE mechanism with platform devices, as the core
> driver code uses the .start and .end to register with the relevant
> iomem reservation code.

Agreed.

> I'm going more strongly on removing the #2 case unless anyone can come up
> with a really compelling reason to keep it.

Ok. I'll submit a patch.

BTW, why do the some boards declare the data resource as more than 4 bytes 
long ?

> > I have DM9000 chips on a custom bus. The bus code remaps the whole address 
> > range (32 bytes per card x 32 slots) in one go and initializes the
> > resource with a subset of the remapped range.
> >
> > Converting that to per-slot ioremaps would create many more mappings. Is
> > that an issue (performance-wise, or in term of a maximum number of
> > mappings if there is a limit) ? I will still need a global mapping, as the
> > bus code needs to access configuration registers on the individual cards,
> > so I won't be able to request_mem_region both the global mapping and the
> > individual mappings.
> > I'm not sure if that's an issue. 
> 
> That sounds like a problem with the bus code mapping too much of the memory
> and thus causing a problem for any subsequent drivers.

That's right. It had been done that way to save mappings. Without a global 
mapping I'll need 64 mappings instead of 1.

> Having an extra mapping in is only a minor problem compared with the other
> travesty the 2 region case is currently causing.

What about 63 extra mappings ?

-- 
Laurent Pinchart
CSE Semaphore Belgium

Chaussee de Bruxelles, 732A
B-1410 Waterloo
Belgium

T +32 (2) 387 42 59
F +32 (2) 387 42 75

Content of type "application/pgp-signature" skipped

Powered by blists - more mailing lists